nanoGPT

nanoGPT is an open-source project that brings GPT-style language modeling directly into spreadsheets, allowing users to experiment with transformer architectures using familiar spreadsheet operations. Built on top of standard spreadsheet functions and formulas, it demonstrates how core components of modern large language models—such as token embeddings, attention mechanisms, and feedforward layers—can be represented, visualized, and manipulated in a tabular environment. Instead of requiring specialized deep learning frameworks, nanoGPT leverages the spreadsheet grid to make each computation step transparent and inspectable.

Key capabilities include defining and adjusting model parameters in cells, tracing data flow across layers, and observing intermediate activations in real time. Users can prototype tiny language models, explore how attention distributes over tokens, and understand how changes in weights affect outputs without writing extensive code. This makes nanoGPT particularly useful for education, research demonstrations, and debugging conceptual understanding of transformers. Typical use cases include teaching the internals of GPT-like models, building minimal working examples of sequence modeling, and experimenting with simplified training or inference workflows. By embedding transformer logic into spreadsheets, nanoGPT lowers the barrier to understanding and experimenting with modern language models while keeping all computations explicit and auditable.